Backend API

As it is currently setup, user code at toolchain blocks running on the platform consumes data stored on disk and produces data which is stored on disk so that subsequent code running on the following blocks can operate on the same principles. This approach allows users to potentially configure experiments with a hybrid set of algorithms that execute on different backends. Each backend can be implemented in a different programming language and contain any number of (pre-installed) libraries users can call on their algorithms.

The requirements for BEAT when reading/writing data are:

  • Ability to manage large and complex data

  • Portability to allow the use of heterogeneous environments

Based on our experience and on these requirements, we investigated the use of HDF5. Unfortunately, HDF5 is not convenient to handle structures such as arrays of variable-size elements, for instance, array of strings. Therefore, we decided to rely on our own binary format.

This document describes the binary formats in BEAT and the API required by BEAT to handle multiple backend implementations. The package beat.env.python27 provides the reference Python backend implementation based on Python 2.7.

Binary Format

Our binary format does not contains information about the format of the data itself, and it is hence necessary to know this format a priori. This means that the format cannot be inferred from the content of a file.

We rely on the following fundamental C-style formats:

  • int8

  • int16

  • int32

  • int64

  • uint8

  • uint16

  • uint32

  • uint64

  • float32

  • float64

  • complex64 (first real value, and then imaginary value)

  • complex128 (first real value, and then imaginary value)

  • bool (written as a byte)

  • string

An element of such a basic format is written in the C-style way, using little-endian byte ordering.

Besides, dataformats always consist of arrays or dictionary of such fundamental formats or compound formats.

An array of elements is saved as followed. First, the shape of the array is saved using an uint64 value for each dimension. Next, the elements of the arrays are saved in C-style order.

A dictionary of elements is saved as followed. First, the key are ordered according to the lexicographic ordering. Then, the values associated to each of these keys are saved following this ordering.

The platform is data-driven and always processes chunks of data. Therefore, data are always written by chunks, each chunk being preceded by a text-formated header indicated the start- and end- indices followed by the size (in bytes) of the chunck.

Considering the Python backend of the platform, this binary format has been successfully implemented using the struct module.

Filesystem Organization

At the filesystem level, each backend shall be organized so it is fully contained in a single-rooted directory tree. The backend installer should not make assumptions about the directory structure of the target operating systems, except, possibly, for the use of stock files. For example:

$ ls /path/to/beat.env.python27
-rw-r--r-- 1 beat beat  2829 Jul 20 19:57 LICENSE
-rw-r--r-- 1 beat beat  7268 Jul 20 19:57 Makefile
-rw-r--r-- 1 beat beat  1852 Jul 20 19:57 README.md
drwxr-xr-x 2 beat beat  4096 Jul 28 16:19 bin/
drwxr-xr-x 2 beat beat  4096 Jul 28 16:19 usr/
drwxr-xr-x 2 beat beat  4096 Jul 28 16:19 src/
...

There is a minimal set of required files in each environment:

  1. Makefile: A make file or equivalent script should be present to fully build the environment from scratch. This allows BEAT platform administrators to install the environment on a target machine.

  2. bin/describe: This is an executable that takes no arguments and describes the current environment, providing its name, version and a list of pre-installed libraries, toolboxes or any other information that may be relevant to users implementing algorithms for this backend. The output of the describe command should be a parseable JSON string. For example, our reference beat.env.python27 environment returns the following for a call to bin/describe:

    {
  "name": "Scientific Python 2.7",
  "version": "0.0.4",
  "os": [
    "Linux",
    "extatix03",
    "3.12.14-1-idiap-generic",
    "#20140313 SMP Thu Mar 13 15:12:40 CET 2014",
    "x86_64",
    ""
  ],
  "packages": {
    "beat.core": "0.9.4",
    "bob": "1.2.2",
    "matplotlib": "1.4.3",
    "numpy": "1.9.2",
    "oset": "0.1.3"
  }
}

    Each pair of name and version for an environment must be unique so that platform users can uniquely select them.

  3. bin/execute: This is an executable that is called by the BEAT infrastructure to execute user code. The executable must be able to receive 2 arguments that correspond to a I/O server address and (temporary) directory containing the following files:

    $ ls -1 /tmp/beat.A976xy1/
configuration.json  #the configuration for the algorithm in JSON format
prefix              #the prefix with algorithm/libraries/formats required

    Optional flags may be provided for administrative purposes, but will not be using for running user code. For example, the reference implementation of bin/execute responds this way when passed the -h optional flag:

    $ /path/to/beat.env.python27/bin/execute -h
Executes a single algorithm.

usage:
  execute [--debug] <addr> <dir>
  execute (--help)


arguments:
  <addr>  Address of the server for I/O requests
  <dir>   Directory containing all configuration required to run the user
          algorithm


options:
  -h, --help   Shows this help message and exit
  -d, --debug  Runs executor in debugging mode

    You should strongly consider implemeting similar functionality on your backend to ease debugging in case of problems.

Further to those files, it is prudent to include:

  1. README.rst: a file containing installation and management instructions. It should preferrably be written using a markup language such as MarkDown (.md extension) or reStructuredText (.rst extension). By reading this file, it should be possible for a remote party with a working knowledge of the target operating system, to completely install the environment without external help.

    The README should also include contact points and, if possible, a bug-tracking link where users can submit bug/update requests.

  2. LICENSE: a file that describes the usage license for the backend.

Message Passing

The BEAT infrastructure communicates with the bin/execute process via Zero Message Queue or ZMQ for short. ZMQ provides a portable bidirectional communication layer between the BEAT infrastructure and the target backend, with many language bindings, including python bindings.

The user process, which manages the data readout of a given algorithm, sends commands back to the infrastructure for requesting data when needed.

"command"
"argument1"
"..."
"argumentn"

Where command is the command name to be executed and argument* are the corresponding arguments, sent using separate zmq.send() calls using the multipart sending technique (as with zmq.SNDMORE). In order to simplify representation, we denote multi-message commands in a single line. So, the command above will be represented in this document such as:

"command" "argument1" "..." "argumentn"

Commands may also exchange binary data. In such a case, we represent it in this manual using <binary>. The binary data format is the one defined by our BaseFormat class at the package beat.backend.python.

The next diagram represents some possible states between the BEAT infrastructure and the execute process in case of a successful execution:

../../../_images/execute.svg

Fig. 9 Message Sequence Chart between BEAT agents and user containers/algorithms

In the remainder of this section, we describe the various commands, which are supported by this communication protocol.

Command: information (ifo)

(User Process -> Infrastructure)

This command asks the infrastructure to return information about the remote data sources queried. The format of this command is:

"ifo name\n"

The infrastructure will answer by writing the following into the input pipe.

"X"
"Start0"
"End0"
...
"StartX-1"
"EndX-1"

where X is the length of the data source and the StartX and EndX are the start and end indexes available through that data source.

Command: get data (get)

(User Process -> Infrastructure)

This command asks the infrastructure to return the data at the given index. The format of this command is:

"get X"

where X is the index of the data in the data source.

The infrastructure will answer by writing the following into the input pipe.

"StartX"
"EndX"
"data"

where StartX and EndX and the start and end indexes in the data sources and data is the packed data that the data sources provides.

Command: done (don)

(User Process -> Infrastructure)

This command notifies the infrastructure that the execution of the user process has completed (i.e. it will not read or write any further data).

"done float"

Where float is a valid floating point number that represents the time wasted waiting for I/O on the user process. This number composes the statistics for processes.

Once the UP has sent this command, the infrastructure will retrieve the statistics (I/O, CPU and memory) and it will acknowlegde the UP, once this is done with:

"ack"

At this point, the UP is expected to gracefully terminate.

Command: error (err)

(User Process -> Infrastructure)

This command notifies the infrastructure that the execution of the user process has err and will not request any further data. A message explaining the error condition is attached.

"err type message"

Once the UP has sent this command, the infrastructure will retrieve the statistics (I/O, CPU and memory) and it will acknowlegde the UP, once this is done with:

"ack"

At this point, the UP is expected to gracefully terminate. The value for type maybe set to usr, indicating the error occurred inside the user code or anything else, indicating it was a system error (and must be reported to system administrators). In this case, the user only gets a generic message indicating a problem with the infrastructure was detected and that system administrators were informed.